01 / Introduction

Samridha
Shrestha

I build segmentation models for maps

AI/ML engineer from Kathmandu. I train deep learning models on geospatial data and build retrieval systems that put language models into real products, along with the web stack that ships them.

See projects Resume (PDF) My apps ↗
Samridha Shrestha

Software engineer at Octacore Solutions

In numbers
Landfill suitability model 96.97%

Accuracy of the custom U-Net, with 0.92 mean IoU and 0.96 Dice.

Flood susceptibility model 95%

CNN accuracy on Kathmandu Valley data, ahead of the Random Forest baseline.

Segmentation architectures compared 3

U-Net, ResU-Net and DeepLabV3, trained and benchmarked on the same geospatial data.

What I do

Three kinds of work

01

Deep learning on maps

U-Net, ResU-Net and DeepLabV3 models that read satellite and terrain layers to find landfill sites and flood risk.

See the research → 02

RAG and LLM systems

Hybrid BM25 and vector search with a reranker and memory, so a shop's chatbot answers with the right product.

See the work → 03

Apps and web

Android apps on Google Play, Wear OS watch faces, and marketing sites for clients, from design to deploy.

apps.samridhashrestha.com.np ↗
Also on Google Play

Small apps that do one thing well

StayT blocks distractions until your task is done. GTA 6 Guide is an offline companion for the game. Anime Studio, a living Wear OS watch face, is next.

  • StayT
  • GTA 6 Guide
  • Anime Studio
Visit apps.samridhashrestha.com.np
apps.samridhashrestha.com.np homepage
Currently

Building retrieval chatbots at Octacore. Keyword search and vector search run side by side, a reranker sorts the results, and the customer gets the right product back.